How to write AI image prompts: describe subject, scene, composition, light and style

Suppose you ask for “a beautiful woman in a café.” The result may be attractive, but almost every important visual decision is still open.

Which café? What time of day? Close-up or full body? Looking at the camera or away? Soft window light or neon? Editorial photography or illustration?

A good image prompt is less like a spell and more like a creative brief.

Start with five layers

A practical structure is:

1. Subject — who or what is the main focus?
2. Scene — where and when is it happening?
3. Composition — how is the image framed?
4. Lighting — where does the light come from and what mood does it create?
5. Style and medium — photo, watercolor, 3D render, fashion editorial, technical diagram and so on.

For example:

A woman reading alone in a quiet Tokyo café at dusk, medium shot from slightly below eye level, rain on the window behind her, soft warm practical lights, natural editorial photography.

This is not guaranteed to produce the perfect result, but it removes a large amount of ambiguity.

Describe visible things, not only emotions

“Lonely” is abstract. “One person at a corner table, empty chairs around her, looking through a rain-covered window” gives the model visible evidence for loneliness.

Likewise, instead of only saying “luxurious,” describe materials, spacing, lighting and objects that communicate luxury.

The general rule is: translate an idea into things that could appear in the frame.

Camera language can control composition

You do not need to be a professional photographer, but a few terms are useful:

• close-up,
• medium shot,
• full-body shot,
• wide shot,
• eye level,
• low angle,
• top-down,
• shallow depth of field,
• centered or asymmetrical composition.

These terms tell the model where visual attention should go.

More words are not always better

A prompt can become self-contradictory. “Minimalist, highly cluttered, soft flat light, dramatic hard spotlight” gives competing directions.

Prioritize what matters most. If one detail is essential, state it clearly rather than hiding it inside a paragraph of decorative adjectives.

Use constraints when the tool supports them

Some image tools support negative prompts, reference images, aspect ratios, style references or edit masks. These controls are often more reliable than repeatedly adding “do not…” to natural-language prompts.

The exact options differ by product, so treat prompt text as one control channel, not the entire system.

Iterate by diagnosing the error

If the subject is right but the framing is wrong, change composition—not the whole prompt.

If the framing is right but the mood is flat, change lighting and color language.

If the character drifts between generations, that is a consistency problem rather than a basic prompting problem. Lesson 015 focuses on that next.

A reusable prompt skeleton

[subject] + [scene] + [composition/camera] + [lighting] + [style/medium] + [important constraints]

You do not need to fill every slot every time. The structure simply helps you notice what you have not specified.

One thing to remember

A useful image prompt describes visual decisions clearly; it does not win by being as long as possible.